[AMD] Add MiniMax-M2.7 accuracy and performance nightly tests (#22722)
Co-authored-by: HaiShaw <hixiao@gmail.com>
This commit is contained in:
Michael
@@ -0,0 +1,245 @@
|
||||
"""AMD MiniMax-M2.7 GSM8K Completion Evaluation Test (8-GPU)
|
||||
|
||||
Tests MiniMax-M2.7 with TP=8 + EP=8 configuration using few-shot completion
|
||||
benchmark on MI325/MI300X.
|
||||
|
||||
Registry: nightly-amd-accuracy-8-gpu-minimax-m27 suite
|
||||
"""
|
||||
|
||||
import ast
|
||||
import os
|
||||
import re
|
||||
import time
|
||||
import unittest
|
||||
from dataclasses import dataclass
|
||||
from typing import List, Optional, Tuple
|
||||
|
||||
import numpy as np
|
||||
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.test.ci.ci_register import register_amd_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
is_in_ci,
|
||||
popen_launch_server,
|
||||
write_github_step_summary,
|
||||
)
|
||||
from sglang.utils import download_and_cache_file, read_jsonl
|
||||
|
||||
register_amd_ci(
|
||||
est_time=3600,
|
||||
suite="nightly-amd-accuracy-8-gpu-minimax-m27",
|
||||
nightly=True,
|
||||
)
|
||||
|
||||
INVALID = -9999999
|
||||
|
||||
|
||||
@dataclass
|
||||
class ModelConfig:
|
||||
"""Configuration for a model to test."""
|
||||
|
||||
model_path: str
|
||||
tp_size: int = 8
|
||||
accuracy_threshold: float = 0.50
|
||||
other_args: Optional[List[str]] = None
|
||||
env_vars: Optional[dict] = None
|
||||
timeout: Optional[int] = None
|
||||
variant: Optional[str] = None
|
||||
|
||||
def __post_init__(self):
|
||||
if self.other_args is None:
|
||||
self.other_args = []
|
||||
if self.env_vars is None:
|
||||
self.env_vars = {}
|
||||
|
||||
def get_display_name(self) -> str:
|
||||
if self.variant:
|
||||
return f"{self.model_path} ({self.variant})"
|
||||
return self.model_path
|
||||
|
||||
|
||||
MINIMAX_M27_MODELS = [
|
||||
ModelConfig(
|
||||
model_path="MiniMaxAI/MiniMax-M2.7",
|
||||
tp_size=8,
|
||||
accuracy_threshold=0.93,
|
||||
timeout=3600,
|
||||
variant="TP8+EP8",
|
||||
other_args=[
|
||||
"--ep-size",
|
||||
"8",
|
||||
"--trust-remote-code",
|
||||
"--attention-backend",
|
||||
"aiter",
|
||||
"--mem-fraction-static",
|
||||
"0.85",
|
||||
"--model-loader-extra-config",
|
||||
'{"enable_multithread_load": true}',
|
||||
"--watchdog-timeout",
|
||||
"1200",
|
||||
],
|
||||
env_vars={"SGLANG_USE_AITER": "1"},
|
||||
),
|
||||
]
|
||||
|
||||
|
||||
def get_one_example(lines, i, include_answer):
|
||||
"""Format a single GSM8K example."""
|
||||
ret = "Question: " + lines[i]["question"] + "\nAnswer:"
|
||||
if include_answer:
|
||||
ret += " " + lines[i]["answer"]
|
||||
return ret
|
||||
|
||||
|
||||
def get_few_shot_examples(lines, k):
|
||||
"""Get k few-shot examples for prompting."""
|
||||
ret = ""
|
||||
for i in range(k):
|
||||
ret += get_one_example(lines, i, True) + "\n\n"
|
||||
return ret
|
||||
|
||||
|
||||
def get_answer_value(answer_str):
|
||||
"""Extract numerical answer from response."""
|
||||
answer_str = answer_str.replace(",", "")
|
||||
numbers = re.findall(r"\d+", answer_str)
|
||||
if len(numbers) < 1:
|
||||
return INVALID
|
||||
try:
|
||||
return ast.literal_eval(numbers[-1])
|
||||
except SyntaxError:
|
||||
return INVALID
|
||||
|
||||
|
||||
def run_gsm8k_benchmark(
|
||||
base_url: str,
|
||||
num_questions: int = 200,
|
||||
num_shots: int = 5,
|
||||
parallel: int = 64,
|
||||
) -> Tuple[float, float, float]:
|
||||
"""Run GSM8K few-shot completion benchmark."""
|
||||
import sglang as sgl
|
||||
from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
|
||||
|
||||
url = "https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl"
|
||||
data_path = download_and_cache_file(url)
|
||||
lines = list(read_jsonl(data_path))
|
||||
|
||||
few_shot_examples = get_few_shot_examples(lines, num_shots)
|
||||
|
||||
questions = []
|
||||
labels = []
|
||||
for i in range(len(lines[:num_questions])):
|
||||
questions.append(get_one_example(lines, i, False))
|
||||
labels.append(get_answer_value(lines[i]["answer"]))
|
||||
assert all(l != INVALID for l in labels)
|
||||
arguments = [{"question": q} for q in questions]
|
||||
|
||||
@sgl.function
|
||||
def few_shot_gsm8k(s, question):
|
||||
s += few_shot_examples + question
|
||||
s += sgl.gen(
|
||||
"answer", max_tokens=512, stop=["Question", "Assistant:", "<|separator|>"]
|
||||
)
|
||||
|
||||
backend = RuntimeEndpoint(base_url)
|
||||
sgl.set_default_backend(backend)
|
||||
|
||||
tic = time.perf_counter()
|
||||
states = few_shot_gsm8k.run_batch(
|
||||
arguments, temperature=0, num_threads=parallel, progress_bar=True
|
||||
)
|
||||
latency = time.perf_counter() - tic
|
||||
|
||||
preds = [get_answer_value(states[i]["answer"]) for i in range(len(states))]
|
||||
acc = np.mean(np.array(preds) == np.array(labels))
|
||||
invalid = np.mean(np.array(preds) == INVALID)
|
||||
|
||||
return float(acc), float(invalid), float(latency)
|
||||
|
||||
|
||||
class TestMiniMaxM27EvalAMD(unittest.TestCase):
|
||||
"""MiniMax-M2.7 GSM8K Completion Evaluation Test for AMD MI325/MI300X."""
|
||||
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.models = MINIMAX_M27_MODELS
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
|
||||
|
||||
def test_minimax_m27_accuracy(self):
|
||||
"""Test MiniMax-M2.7 with GSM8K completion benchmark."""
|
||||
all_results = []
|
||||
summary = "### MiniMax-M2.7 Models (MI325)\n\n"
|
||||
summary += "| Model | Variant | TP | Accuracy | Threshold | Status |\n"
|
||||
summary += "| ----- | ------- | -- | -------- | --------- | ------ |\n"
|
||||
|
||||
for config in self.models:
|
||||
display_name = config.get_display_name()
|
||||
with self.subTest(model=display_name):
|
||||
print(f"\n{'='*60}")
|
||||
print(f"Testing: {display_name}")
|
||||
print(f"{'='*60}")
|
||||
|
||||
env = os.environ.copy()
|
||||
for key, value in config.env_vars.items():
|
||||
env[key] = value
|
||||
|
||||
other_args = list(config.other_args)
|
||||
other_args.extend(["--tp", str(config.tp_size)])
|
||||
timeout = config.timeout or DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
|
||||
|
||||
try:
|
||||
process = popen_launch_server(
|
||||
model=config.model_path,
|
||||
base_url=self.base_url,
|
||||
timeout=timeout,
|
||||
other_args=other_args,
|
||||
env=env,
|
||||
)
|
||||
|
||||
try:
|
||||
acc, invalid, latency = run_gsm8k_benchmark(
|
||||
self.base_url, num_questions=self.num_questions
|
||||
)
|
||||
passed = acc >= config.accuracy_threshold
|
||||
status = "PASS" if passed else "FAIL"
|
||||
print(
|
||||
f" accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
|
||||
)
|
||||
|
||||
all_results.append(
|
||||
{
|
||||
"model": display_name,
|
||||
"accuracy": acc,
|
||||
"passed": passed,
|
||||
}
|
||||
)
|
||||
summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
|
||||
|
||||
finally:
|
||||
kill_process_tree(process.pid)
|
||||
|
||||
except Exception as e:
|
||||
summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | N/A | {config.accuracy_threshold} | ERROR |\n"
|
||||
all_results.append(
|
||||
{
|
||||
"model": display_name,
|
||||
"accuracy": None,
|
||||
"passed": False,
|
||||
"error": str(e),
|
||||
}
|
||||
)
|
||||
|
||||
if is_in_ci():
|
||||
write_github_step_summary(summary)
|
||||
|
||||
failed = [r for r in all_results if not r["passed"]]
|
||||
if failed:
|
||||
raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,249 @@
|
||||
"""MI35x MiniMax-M2.7 GSM8K Completion Evaluation Test (8-GPU)
|
||||
|
||||
Tests MiniMax-M2.7 with TP=8 + EP=8 configuration using few-shot completion
|
||||
benchmark on MI35x.
|
||||
|
||||
Registry: nightly-amd-8-gpu-mi35x-minimax-m27 suite
|
||||
"""
|
||||
|
||||
import ast
|
||||
import os
|
||||
|
||||
os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
|
||||
os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
|
||||
|
||||
import re
|
||||
import time
|
||||
import unittest
|
||||
from dataclasses import dataclass
|
||||
from typing import List, Optional, Tuple
|
||||
|
||||
import numpy as np
|
||||
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.test.ci.ci_register import register_amd_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
is_in_ci,
|
||||
popen_launch_server,
|
||||
write_github_step_summary,
|
||||
)
|
||||
from sglang.utils import download_and_cache_file, read_jsonl
|
||||
|
||||
register_amd_ci(
|
||||
est_time=5400,
|
||||
suite="nightly-amd-8-gpu-mi35x-minimax-m27",
|
||||
nightly=True,
|
||||
)
|
||||
|
||||
INVALID = -9999999
|
||||
|
||||
|
||||
@dataclass
|
||||
class ModelConfig:
|
||||
"""Configuration for a model to test."""
|
||||
|
||||
model_path: str
|
||||
tp_size: int = 8
|
||||
accuracy_threshold: float = 0.50
|
||||
other_args: Optional[List[str]] = None
|
||||
env_vars: Optional[dict] = None
|
||||
timeout: Optional[int] = None
|
||||
variant: Optional[str] = None
|
||||
|
||||
def __post_init__(self):
|
||||
if self.other_args is None:
|
||||
self.other_args = []
|
||||
if self.env_vars is None:
|
||||
self.env_vars = {}
|
||||
|
||||
def get_display_name(self) -> str:
|
||||
if self.variant:
|
||||
return f"{self.model_path} ({self.variant})"
|
||||
return self.model_path
|
||||
|
||||
|
||||
MI35X_MINIMAX_M27_MODELS = [
|
||||
ModelConfig(
|
||||
model_path="MiniMaxAI/MiniMax-M2.7",
|
||||
tp_size=8,
|
||||
accuracy_threshold=0.93,
|
||||
timeout=5400,
|
||||
variant="TP8+EP8",
|
||||
other_args=[
|
||||
"--ep-size",
|
||||
"8",
|
||||
"--trust-remote-code",
|
||||
"--attention-backend",
|
||||
"aiter",
|
||||
"--mem-fraction-static",
|
||||
"0.85",
|
||||
"--model-loader-extra-config",
|
||||
'{"enable_multithread_load": true}',
|
||||
"--watchdog-timeout",
|
||||
"1200",
|
||||
],
|
||||
env_vars={"SGLANG_USE_AITER": "1"},
|
||||
),
|
||||
]
|
||||
|
||||
|
||||
def get_one_example(lines, i, include_answer):
|
||||
"""Format a single GSM8K example."""
|
||||
ret = "Question: " + lines[i]["question"] + "\nAnswer:"
|
||||
if include_answer:
|
||||
ret += " " + lines[i]["answer"]
|
||||
return ret
|
||||
|
||||
|
||||
def get_few_shot_examples(lines, k):
|
||||
"""Get k few-shot examples for prompting."""
|
||||
ret = ""
|
||||
for i in range(k):
|
||||
ret += get_one_example(lines, i, True) + "\n\n"
|
||||
return ret
|
||||
|
||||
|
||||
def get_answer_value(answer_str):
|
||||
"""Extract numerical answer from response."""
|
||||
answer_str = answer_str.replace(",", "")
|
||||
numbers = re.findall(r"\d+", answer_str)
|
||||
if len(numbers) < 1:
|
||||
return INVALID
|
||||
try:
|
||||
return ast.literal_eval(numbers[-1])
|
||||
except SyntaxError:
|
||||
return INVALID
|
||||
|
||||
|
||||
def run_gsm8k_benchmark(
|
||||
base_url: str,
|
||||
num_questions: int = 200,
|
||||
num_shots: int = 5,
|
||||
parallel: int = 64,
|
||||
) -> Tuple[float, float, float]:
|
||||
"""Run GSM8K few-shot completion benchmark."""
|
||||
import sglang as sgl
|
||||
from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
|
||||
|
||||
url = "https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl"
|
||||
data_path = download_and_cache_file(url)
|
||||
lines = list(read_jsonl(data_path))
|
||||
|
||||
few_shot_examples = get_few_shot_examples(lines, num_shots)
|
||||
|
||||
questions = []
|
||||
labels = []
|
||||
for i in range(len(lines[:num_questions])):
|
||||
questions.append(get_one_example(lines, i, False))
|
||||
labels.append(get_answer_value(lines[i]["answer"]))
|
||||
assert all(l != INVALID for l in labels)
|
||||
arguments = [{"question": q} for q in questions]
|
||||
|
||||
@sgl.function
|
||||
def few_shot_gsm8k(s, question):
|
||||
s += few_shot_examples + question
|
||||
s += sgl.gen(
|
||||
"answer", max_tokens=512, stop=["Question", "Assistant:", "<|separator|>"]
|
||||
)
|
||||
|
||||
backend = RuntimeEndpoint(base_url)
|
||||
sgl.set_default_backend(backend)
|
||||
|
||||
tic = time.perf_counter()
|
||||
states = few_shot_gsm8k.run_batch(
|
||||
arguments, temperature=0, num_threads=parallel, progress_bar=True
|
||||
)
|
||||
latency = time.perf_counter() - tic
|
||||
|
||||
preds = [get_answer_value(states[i]["answer"]) for i in range(len(states))]
|
||||
acc = np.mean(np.array(preds) == np.array(labels))
|
||||
invalid = np.mean(np.array(preds) == INVALID)
|
||||
|
||||
return float(acc), float(invalid), float(latency)
|
||||
|
||||
|
||||
class TestMiniMaxM27EvalMI35x(unittest.TestCase):
|
||||
"""MiniMax-M2.7 GSM8K Completion Evaluation Test for AMD MI35x."""
|
||||
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.models = MI35X_MINIMAX_M27_MODELS
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
|
||||
|
||||
def test_minimax_m27_accuracy(self):
|
||||
"""Test MiniMax-M2.7 with GSM8K completion benchmark."""
|
||||
all_results = []
|
||||
summary = "### MiniMax-M2.7 Models (MI35x)\n\n"
|
||||
summary += "| Model | Variant | TP | Accuracy | Threshold | Status |\n"
|
||||
summary += "| ----- | ------- | -- | -------- | --------- | ------ |\n"
|
||||
|
||||
for config in self.models:
|
||||
display_name = config.get_display_name()
|
||||
with self.subTest(model=display_name):
|
||||
print(f"\n{'='*60}")
|
||||
print(f"Testing: {display_name}")
|
||||
print(f"{'='*60}")
|
||||
|
||||
env = os.environ.copy()
|
||||
for key, value in config.env_vars.items():
|
||||
env[key] = value
|
||||
|
||||
other_args = list(config.other_args)
|
||||
other_args.extend(["--tp", str(config.tp_size)])
|
||||
timeout = config.timeout or DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
|
||||
|
||||
try:
|
||||
process = popen_launch_server(
|
||||
model=config.model_path,
|
||||
base_url=self.base_url,
|
||||
timeout=timeout,
|
||||
other_args=other_args,
|
||||
env=env,
|
||||
)
|
||||
|
||||
try:
|
||||
acc, invalid, latency = run_gsm8k_benchmark(
|
||||
self.base_url, num_questions=self.num_questions
|
||||
)
|
||||
passed = acc >= config.accuracy_threshold
|
||||
status = "PASS" if passed else "FAIL"
|
||||
print(
|
||||
f" accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
|
||||
)
|
||||
|
||||
all_results.append(
|
||||
{
|
||||
"model": display_name,
|
||||
"accuracy": acc,
|
||||
"passed": passed,
|
||||
}
|
||||
)
|
||||
summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
|
||||
|
||||
finally:
|
||||
kill_process_tree(process.pid)
|
||||
|
||||
except Exception as e:
|
||||
summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | N/A | {config.accuracy_threshold} | ERROR |\n"
|
||||
all_results.append(
|
||||
{
|
||||
"model": display_name,
|
||||
"accuracy": None,
|
||||
"passed": False,
|
||||
"error": str(e),
|
||||
}
|
||||
)
|
||||
|
||||
if is_in_ci():
|
||||
write_github_step_summary(summary)
|
||||
|
||||
failed = [r for r in all_results if not r["passed"]]
|
||||
if failed:
|
||||
raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,140 @@
|
||||
"""Nightly performance benchmark for MiniMax-M2.7 on MI325/MI300X (8-GPU).
|
||||
|
||||
This test benchmarks MiniMax-M2.7 with TP=8 + EP=8 configuration.
|
||||
|
||||
The model path can be configured via MINIMAX_M27_MODEL_PATH environment variable.
|
||||
|
||||
Registry: nightly-perf-8-gpu-minimax-m27 suite
|
||||
|
||||
Example usage:
|
||||
python -m pytest test_minimax_m27_perf_amd.py -v
|
||||
"""
|
||||
|
||||
import os
|
||||
import unittest
|
||||
from typing import List
|
||||
|
||||
from sglang.test.ci.ci_register import register_amd_ci
|
||||
from sglang.test.nightly_bench_utils import BenchmarkResult
|
||||
from sglang.test.nightly_utils import NightlyBenchmarkRunner
|
||||
from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
|
||||
|
||||
register_amd_ci(est_time=5400, suite="nightly-perf-8-gpu-minimax-m27", nightly=True)
|
||||
|
||||
|
||||
def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
|
||||
"""Generate a simplified markdown report without traces and cost columns.
|
||||
|
||||
Skips the first result if it's a warmup run (duplicate batch_size).
|
||||
"""
|
||||
model_header = results[0].model_path
|
||||
if results[0].run_name and results[0].run_name != "default":
|
||||
model_header += f" ({results[0].run_name})"
|
||||
|
||||
gpu_config = os.getenv("GPU_CONFIG", "MI325")
|
||||
if gpu_config:
|
||||
model_header += f" [{gpu_config}]"
|
||||
|
||||
summary = f"### {model_header}\n"
|
||||
summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
|
||||
summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
|
||||
|
||||
report_results = (
|
||||
results[1:]
|
||||
if len(results) > 1 and results[0].batch_size == results[1].batch_size
|
||||
else results
|
||||
)
|
||||
|
||||
for result in report_results:
|
||||
itl = 1 / (result.output_throughput / result.batch_size) * 1000
|
||||
summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
|
||||
|
||||
return summary
|
||||
|
||||
|
||||
MINIMAX_M27_MODEL_PATH = os.environ.get(
|
||||
"MINIMAX_M27_MODEL_PATH", "MiniMaxAI/MiniMax-M2.7"
|
||||
)
|
||||
PROFILE_DIR = "performance_profiles_minimax_m27"
|
||||
|
||||
|
||||
class TestNightlyMiniMaxM27Performance(unittest.TestCase):
|
||||
"""Nightly performance benchmark for MiniMax-M2.7 on MI325/MI300X.
|
||||
|
||||
Tests MiniMax-M2.7 with TP=8 + EP=8 configuration.
|
||||
"""
|
||||
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.batch_sizes = [1, 8, 16, 64]
|
||||
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
|
||||
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
|
||||
|
||||
cls.model_config = {
|
||||
"name": "minimax-m27-tp8-ep8",
|
||||
"model_path": MINIMAX_M27_MODEL_PATH,
|
||||
"other_args": [
|
||||
"--trust-remote-code",
|
||||
"--tp",
|
||||
"8",
|
||||
"--ep-size",
|
||||
"8",
|
||||
"--attention-backend",
|
||||
"aiter",
|
||||
"--mem-fraction-static",
|
||||
"0.85",
|
||||
"--model-loader-extra-config",
|
||||
'{"enable_multithread_load": true}',
|
||||
"--watchdog-timeout",
|
||||
"1200",
|
||||
],
|
||||
"env_vars": {
|
||||
"SGLANG_USE_AITER": "1",
|
||||
},
|
||||
}
|
||||
|
||||
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
|
||||
cls.runner.setup_profile_directory()
|
||||
cls.runner.full_report = f"## {cls.__name__}\n"
|
||||
|
||||
def test_bench_minimax_m27(self):
|
||||
"""Run benchmark for MiniMax-M2.7."""
|
||||
old_env = {}
|
||||
for key, value in self.model_config.get("env_vars", {}).items():
|
||||
old_env[key] = os.environ.get(key)
|
||||
os.environ[key] = value
|
||||
print(f"Setting env: {key}={value}")
|
||||
|
||||
try:
|
||||
result_tuple = self.runner.run_benchmark_for_model(
|
||||
model_path=self.model_config["model_path"],
|
||||
batch_sizes=self.batch_sizes,
|
||||
input_lens=self.input_lens,
|
||||
output_lens=self.output_lens,
|
||||
other_args=self.model_config["other_args"],
|
||||
variant=self.model_config["name"],
|
||||
extra_bench_args=["--trust-remote-code"],
|
||||
enable_profile=False,
|
||||
timeout=5400,
|
||||
)
|
||||
results = result_tuple[0]
|
||||
success = result_tuple[1]
|
||||
|
||||
if results:
|
||||
self.runner.full_report += (
|
||||
generate_simple_markdown_report(results) + "\n"
|
||||
)
|
||||
|
||||
self.assertTrue(success, "Benchmark failed for MiniMax-M2.7")
|
||||
finally:
|
||||
for key, value in old_env.items():
|
||||
if value is None:
|
||||
os.environ.pop(key, None)
|
||||
else:
|
||||
os.environ[key] = value
|
||||
self.runner.write_final_report()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,146 @@
|
||||
"""MI35x Nightly performance benchmark for MiniMax-M2.7 (8-GPU).
|
||||
|
||||
This test benchmarks MiniMax-M2.7 with TP=8 + EP=8 configuration on MI35x.
|
||||
|
||||
The model path can be configured via MINIMAX_M27_MODEL_PATH environment variable.
|
||||
|
||||
Registry: nightly-perf-8-gpu-mi35x-minimax-m27 suite
|
||||
|
||||
Example usage:
|
||||
python -m pytest test_minimax_m27_perf_mi35x.py -v
|
||||
"""
|
||||
|
||||
import os
|
||||
|
||||
os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
|
||||
os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
|
||||
|
||||
import unittest
|
||||
from typing import List
|
||||
|
||||
from sglang.test.ci.ci_register import register_amd_ci
|
||||
from sglang.test.nightly_bench_utils import BenchmarkResult
|
||||
from sglang.test.nightly_utils import NightlyBenchmarkRunner
|
||||
from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
|
||||
|
||||
register_amd_ci(
|
||||
est_time=5400, suite="nightly-perf-8-gpu-mi35x-minimax-m27", nightly=True
|
||||
)
|
||||
|
||||
|
||||
def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
|
||||
"""Generate a simplified markdown report without traces and cost columns.
|
||||
|
||||
Skips the first result if it's a warmup run (duplicate batch_size).
|
||||
"""
|
||||
model_header = results[0].model_path
|
||||
if results[0].run_name and results[0].run_name != "default":
|
||||
model_header += f" ({results[0].run_name})"
|
||||
|
||||
gpu_config = os.getenv("GPU_CONFIG", "MI35x")
|
||||
if gpu_config:
|
||||
model_header += f" [{gpu_config}]"
|
||||
|
||||
summary = f"### {model_header}\n"
|
||||
summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
|
||||
summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
|
||||
|
||||
report_results = (
|
||||
results[1:]
|
||||
if len(results) > 1 and results[0].batch_size == results[1].batch_size
|
||||
else results
|
||||
)
|
||||
|
||||
for result in report_results:
|
||||
itl = 1 / (result.output_throughput / result.batch_size) * 1000
|
||||
summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
|
||||
|
||||
return summary
|
||||
|
||||
|
||||
MINIMAX_M27_MODEL_PATH = os.environ.get(
|
||||
"MINIMAX_M27_MODEL_PATH", "MiniMaxAI/MiniMax-M2.7"
|
||||
)
|
||||
PROFILE_DIR = "performance_profiles_minimax_m27_mi35x"
|
||||
|
||||
|
||||
class TestNightlyMiniMaxM27PerformanceMI35x(unittest.TestCase):
|
||||
"""MI35x Nightly performance benchmark for MiniMax-M2.7.
|
||||
|
||||
Tests MiniMax-M2.7 with TP=8 + EP=8 configuration.
|
||||
"""
|
||||
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.batch_sizes = [1, 8, 16, 64]
|
||||
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
|
||||
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
|
||||
|
||||
cls.model_config = {
|
||||
"name": "minimax-m27-tp8-ep8",
|
||||
"model_path": MINIMAX_M27_MODEL_PATH,
|
||||
"other_args": [
|
||||
"--trust-remote-code",
|
||||
"--tp",
|
||||
"8",
|
||||
"--ep-size",
|
||||
"8",
|
||||
"--attention-backend",
|
||||
"aiter",
|
||||
"--mem-fraction-static",
|
||||
"0.85",
|
||||
"--model-loader-extra-config",
|
||||
'{"enable_multithread_load": true}',
|
||||
"--watchdog-timeout",
|
||||
"1200",
|
||||
],
|
||||
"env_vars": {
|
||||
"SGLANG_USE_AITER": "1",
|
||||
},
|
||||
}
|
||||
|
||||
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
|
||||
cls.runner.setup_profile_directory()
|
||||
cls.runner.full_report = f"## {cls.__name__}\n"
|
||||
|
||||
def test_bench_minimax_m27(self):
|
||||
"""Run benchmark for MiniMax-M2.7."""
|
||||
old_env = {}
|
||||
for key, value in self.model_config.get("env_vars", {}).items():
|
||||
old_env[key] = os.environ.get(key)
|
||||
os.environ[key] = value
|
||||
print(f"Setting env: {key}={value}")
|
||||
|
||||
try:
|
||||
result_tuple = self.runner.run_benchmark_for_model(
|
||||
model_path=self.model_config["model_path"],
|
||||
batch_sizes=self.batch_sizes,
|
||||
input_lens=self.input_lens,
|
||||
output_lens=self.output_lens,
|
||||
other_args=self.model_config["other_args"],
|
||||
variant=self.model_config["name"],
|
||||
extra_bench_args=["--trust-remote-code"],
|
||||
enable_profile=False,
|
||||
timeout=5400,
|
||||
)
|
||||
results = result_tuple[0]
|
||||
success = result_tuple[1]
|
||||
|
||||
if results:
|
||||
self.runner.full_report += (
|
||||
generate_simple_markdown_report(results) + "\n"
|
||||
)
|
||||
|
||||
self.assertTrue(success, "Benchmark failed for MiniMax-M2.7 on MI35x")
|
||||
finally:
|
||||
for key, value in old_env.items():
|
||||
if value is None:
|
||||
os.environ.pop(key, None)
|
||||
else:
|
||||
os.environ[key] = value
|
||||
self.runner.write_final_report()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user